Explainable machine learning for predicting thermogravimetric analysis of oxidatively torrefied spent coffee grounds combustion

Understanding the combustion behavior of oxidatively torrefied spent coffee grounds (SCG) is crucial for advancing sustainable fuel technologies. This study introduces a novel, explainable machine learning (ML) framework as a cost-effective alternative to traditional thermogravimetric analysis (TGA) that is designed to accelerate the evaluation of oxidatively torrefied SCG combustion properties. Four ML models: artificial neural network (ANN), k-nearest neighbor (k-NN), random forest (RF), and decision tree (DT), were compared to predict TGA data using proximate analysis and combustion temperature (CT). Among the evaluated models, k-NN exhibited the highest performance, achieving near-perfect R² values that exceeded 0.9904 and RMSE values below 0.9552 on the validation set for both TG (mass loss) and DTG (derivative mass loss). It also accurately predicted key combustion properties, including ignition, peak, and burnout temperature when tested on unknown data. LIME (Local Interpretable Model-agnostic Explanations) analysis revealed that CT was the most influential predictor for TG and DTG, enhancing model interpretability. The results highlight the effectiveness of the k-NN-LIME approach in analyzing the combustion of oxidatively torrefied SCG, offering a robust and explainable model with significant implications for bioenergy research and sustainable fuel development.